The MÉTRON platform

The founder-donated no/low-code MÉTRON platform is the project's most substantial in-kind contribution. The platform is the frontend that exposes the multi-language MÉTRON model family to non-ML researchers (philosophers of language, theologians, computational linguists, cognitive psychologists, cross-tradition scholars), letting them run controlled cross-linguistic transformer experiments themselves at discretionary spending costs, without ML-engineering skill.

What the platform exposes: the measurement levers

The platform does not ask a researcher to train a model or read its internals. It exposes a small set of measurement levers, and the machine learning stays out of sight. The program's results to date generalize into four reusable operations:

• The ablation lever. Remove or vary a single feature of a language, train matched models on corpora that differ in exactly that feature, and measure what the change causally shapes. This is the controlled experiment on language structure that cannot be run on human beings.
• The alignment lever. Fit a single map between two meaning spaces (two languages, two traditions, two periods) and measure how much of one is recoverable from the other. What aligns is the shared core; what resists is the part bound to one side. Translatability and universality become measurements rather than intuitions.
• The structure lever. Read the deep statistical signature of a body of text: whether it carries the structure associated with competence, with oral composition, with literariness, or with a distinct authorial hand.
• The small-corpus condition. Competence tracks the structural richness of the input, not its sheer volume. A dead language, a single author, a single period, a single tradition, each far too small for web-scale methods, becomes a workable subject.

A research question becomes an experiment by being fitted to one or more of these levers, with a falsifier stated in advance and a placebo arm attached. The platform is small because the levers are few, and its reach is large because the questions that reduce to a configuration of these four are many.

What the telescope makes tractable

Why should a language-model experiment bear on a question in philosophy or history at all? Because the knowledge those questions reach for is deposited in language itself. Consider the sentence "I declare you man and wife." Whether it succeeds has nothing to do with grammar; it depends on who says it, to whom, and under what authority. All of that is carried in language and its settings, and a competent reader knows it from having read enough language. That is the project's central claim carried past grammar: the structure latent in language holds far more than the rules of syntax, and an instrument trained only on language can measure where that structure is shared and where it is bound.

So the levers open questions that were long held to be matters of interpretation alone:

• Is a concept universal or culture-bound? The alignment lever measures whether a moral, mystical, or metaphysical vocabulary aligns across traditions or resolves only within each one.
• Did a concept change meaning across an era? The same lever, run across periods, turns conceptual history and questions of legal originalism into a measurement.
• Does the grammar of a language shape what its speakers can readily express? The ablation lever makes the old linguistic-relativity question causal rather than correlational.
• Whose hand wrote a text, and was it composed orally or in writing? The structure lever reads the signature that separates authors and composition modes.
• How do two civilizational traditions compare when neither left a web-scale corpus? The small-corpus condition makes a single philosophical canon or a single scriptural tradition a viable subject.

The instrument settles the descriptive question, whether a structure is shared or bound, present or absent, and never the normative one, whether a morality is correct or a tradition true. That boundary is the discipline the whole project is built on.

Infrastructure

The platform runs on a single NVIDIA H100 80GB GPU acquired once funded as grant-funded capital infrastructure (chassis, networking, and one-year hosting included). The H100 is the grant's persistent skin in the game: it stays with the project past grant end and hosts the platform indefinitely. A community-contribution channel offers free compute and founder support to native-speaker linguists wishing to develop minimal-pair benchmarks for their own languages, in exchange for contributing the resulting benchmarks back to the open MÉTRON family. Each contributed benchmark expands what the platform can ask, in the contributor's native language, without further grant funding.

Two-stage public trajectory

The platform follows a deliberate two-stage public trajectory:

• An early proof-of-concept will be announced at EMNLP 2026 (system demonstrations track, submission August-September 2026).
• The working platform will launch at EMNLP 2027 with funder acknowledgment.

Two consecutive years of visible presence at the field's primary venue, with the second year carrying the funder's name on a working artifact actively being adopted by the community.

What it costs

Per-language costs vary substantially with morphology, tokenizer training, corpus availability, and ablation depth. French came in at $65; others will run higher or lower — but all stay at discretionary spending level costs rather than institutional-cluster budgets. The methodology is itself a structural contribution: it opens empirical access to philosophical and linguistic questions previously gated behind institutional compute. A philosopher of language asking which structural features license which kinds of grammatical generalization, an applied linguist working on under-resourced languages, a graduate student in computational linguistics — each can now run the controlled experiment at consumer-GPU cost. The democratization is the form the work takes.

What it enables

Larger-scale parameter experiments (1B and 5B parameter ablations at multi-language scale) are deliberately Phase 2 work, contingent on the capital and fiscal-sponsorship arrangements that build on this Phase 1 foundation.

Phase 2: 4E cognition empirical track

A second empirical track on 4E cognition (extended, embodied, embedded, enacted), using the multicardz polymorphic spatial-attribute system (the founder's proprietary product) as the experimental vehicle, is Phase 2 work. Three studies are already pre-registered on the Open Science Framework:

• Attribute-Uniqueness Study (OSF XWKP5, N=80)
• Duplicate-Tolerant Matrix Eye-Tracking Study (OSF JZUBN, N=24)
• Bloom-Filter Threshold Study (OSF 2YUNF, N=120)

A philosopher of cognitive science specializing in the 4E framework is the intended Phase 2 collaborator. Phase 1 includes outreach to that collaborator as a relationship-building deliverable; the experiments themselves run in Phase 2 once the collaboration is confirmed. multicardz remains the founder's proprietary product; for the 4E track in Phase 2, the founder will donate free premium accounts (as many as needed, assigned to the project for its duration), associated server compute, and founder time on the 4E line, at zero cost to the funder budget.